---
license: cc-by-nc-sa-4.0
pretty_name: MovingFashion Image-to-Video Retrieval
tags:
- mteb
- moeb
- image-to-video
- cross-modal-retrieval
configs:
- config_name: corpus
  data_files:
  - split: test
    path: corpus/test-*
- config_name: qrels
  data_files:
  - split: test
    path: qrels/test-*
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*
dataset_info:
  config_name: corpus
  features:
  - name: id
    dtype: string
  - name: video
    dtype: video
  splits:
  - name: test
    num_bytes: 2443458284
    num_examples: 1328
  download_size: 2171821800
  dataset_size: 2443458284
---

# MovingFashion Image-to-Video Retrieval

Frozen MTEB/MOEB representation of the official MovingFashion test split for
shop-image-to-video fashion retrieval.

## Construction

The dataset is derived from the [official MovingFashion release](https://humaticslab.github.io/retrieval/movingfashion)
and its `test.json` associations. The MTEB construction script audits both
official splits against the source code at revision
[`4ca15d147ce87f0385c0c9779eac49e55c727ec8`](https://github.com/HumaticsLAB/SEAM-Match-RCNN/tree/4ca15d147ce87f0385c0c9779eac49e55c727ec8),
checks that train and test share no product IDs or media paths, verifies every
annotation reference against the archive, pins the known source omissions, and
decodes all published test media. The source archive SHA-256 for this build is
`20ae89a67a58d3dfc2304c2533d5a5c684eb6888aec9a7d7d3eda22b0a81f5f4`.

Media paths are used as IDs. Repeated paths are collapsed without discarding
associations, so the source's multi-positive relevance structure is preserved.
The Hub configs use the standard MTEB representation: `queries` contains
shop-image queries, `corpus` contains social videos, and `qrels`
contains binary relevance judgments.

The official archive omits 22 train videos and one annotated test video. The
The shop image whose only annotated video is missing is excluded from the query set together with its unusable qrel. The construction script pins and reports all 23 source
omissions rather than silently dropping them.

## Evaluation contents

- shop-image-to-video: 1340 queries,
  1328 corpus items, and 1341 qrels.
- Queries with multiple positives: 1.
- Corpus items without a qrel: 0.
- Source difficulty labels: `0` is hard and `1` is regular. They are audited
  during construction but are not used to filter the benchmark.

## Source protocol and baselines

The source benchmark evaluates video-to-shop retrieval with top-k accuracy. The
paper reports SEAM Match-RCNN top-1/5/10/20 accuracy of .49/.80/.89/.94 overall,
.55/.86/.94/.97 on the regular subset, and .30/.62/.76/.87 on the hard subset.
Those numbers use the original task-specific detector and are context rather
than directly comparable guarantees for generic embedding models.

## License, provenance, and limitations

The official repository labels the work CC BY-NC-SA 4.0 and says the dataset is
available for academic purposes. This derived release therefore retains
CC BY-NC-SA 4.0, attribution, non-commercial, and share-alike requirements.
Source videos originated on Instagram and shop images on Net-A-Porter; underlying
media, privacy, publicity, trademark, and platform rights may remain with their
respective owners. The paper states that faces were blurred. Users remain
responsible for determining whether their use complies with the license,
source-platform terms, and applicable law.

## Citation

```bibtex
@misc{godi2021movingfashion,
  title = {MovingFashion: a Benchmark for the Video-to-Shop Challenge},
  author = {Godi, Marco and Joppi, Christian and Skenderi, Geri and Cristani, Marco},
  year = {2021},
  eprint = {2110.02627},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
}
```

See the [paper](https://arxiv.org/abs/2110.02627) and [official code](https://github.com/HumaticsLAB/SEAM-Match-RCNN).